Tesouro Direto prices re-decided from the official rule table

The field is the result: an almost uniform texture of reproduced prices, with the exceptions where they are. The buy side is drawn because it carries the settlement convention; on the sell side (D+1 until 2021-09-10, D+0 since) 91.16% of 75,982 prices are REPRODUCED, 6,551 are CONSISTENT and 164 OFF, and its exceptions are in the table of §3.

Three things the methodology does not say about Tesouro Direto’s published prices were pre-registered as hypotheses and decided row by row: on which day a purchase and a sale settle, and how a 6-decimal price becomes the 2-decimal one printed. The census answers: buy D+1 across the whole history; sell D+1 until 2021-09-10, sell D+0 since — a change of settlement convention made during 2021 and not printed beside the price (770 sell-side prices of 2021 reproduce only under D+0, every one from 2022 on; 28 earlier prices do too, on 9 days where both settlements were quoted); centavos truncated throughout. Where every hypothesis fails, the row is OFF; the table shows how many, by year.

A calendar is dated. The pinned holiday file is the one ANBIMA serves today, and it carries a holiday that did not exist when most of these prices were published: 20 November, national since Lei 14.759 of 21 December 2023. A quotation of 2019 for a title maturing in 2029 was priced with the calendar of 2019, in which 2024, 2025, 2026 and 2028 each had one more business day than today’s file says (2027’s 20 November is a Saturday). Re-decided with today’s calendar those prices are OFF by exactly that many days; re-decided with the calendar as of the quotation they are the rule’s value to the centavo. The pre-registration’s second amendment added that hypothesis after the full run showed the pattern (du + 1 for 2025 maturities, + 2 for 2026, + 4 for 2029 on a seeded sample of 30 rows); the rows that need it are the heavier bright lines in the atlas and the “of which old calendar only” columns here. Around the law’s date the census reads, per quotation day, how many rows reproduce only under today’s calendar and how many only under the old one; the last day priced on the old calendar is 2023-12-22 and the first priced on the new one is 2023-12-26.

The settlement calendar is the exchange’s. The methodology counts du on the national bank calendar, but Tesouro Direto settles through B3, which did not operate on São Paulo’s own holidays (25 January, 9 July, and 20 November from 2006) through 2021, on 24 and 31 December every year, and on the 2014 World Cup days the exchange listed. A purchase quoted the day before settles on the next day B3 traded, and the published price says so. corpus/pu/b3-settlement.json derives those days from B3’s circulars (2017–2026) and from the exchange’s calendar pages as the Wayback Machine captured them (2004–2016), citing for each date the documents that name the holiday; dates no captured document names (2007 and 2008 among them) follow the rule alone and the prices resting on them are counted apart. The July 2014 match days are not in the table because no page listing them was captured; those prices stay OFF.

The settlement skip: Tesouro Direto quotes on 24 and 31 December but does not settle on them, and the ANBIMA holiday file says in its own notes that it lists only days that close bank reserves and omits the last business day of the year. The amendment of 2026-10-16 in the pre-registration added that hypothesis after the first 4,000 rows showed it; the rows that need it are counted apart, never folded into the primary count.

The first 80 of 289 rows; every row is in certs/pu-ledger.json (offRows). A difference of a few centavos with the sign of a later settlement is a settlement-day convention; a difference of about R$ 48 on an NTN-F quoted on the last business day of the year is a coupon gone ex because settlement crossed 1 January. Of the 289 rows that remain OFF after the B3 settlement calendar, 168 are the rule’s value for a settlement one to six bank business days later than any pre-registered hypothesis — the record carries that implied settlement for each, and it is never used to decide. The largest groups: 2008-06-19 → 2008-06-23 (28), 2014-06-27 → 2014-07-01 (17), 2014-06-25 → 2014-06-27 (17), 2016-12-30 → 2017-01-03 (15), 2016-12-29 → 2017-01-03 (15), 2018-02-01 → 2018-02-05 (14), 2015-12-31 → 2016-01-05 (14), 2015-12-30 → 2016-01-05 (14). Three of those groups are the last quotation days of 2014, 2015 and 2016, whose prices are the value for settlement on the SECOND business day of the next year; the exchange’s circulars state that practice for 2018 and 2019 (OC 111/2016: “Por solicitação da Secretaria do Tesouro Nacional, a primeira liquidação de compra e de venda de 2018 será realizada no segundo dia útil”, and OC 021/2017 the same for 2019), yet the published prices of 28 December 2017 and 28 December 2018 are the value for the first business day. A circular and the published numbers disagree; neither is assumed here, and no first-business-day skip is in the settlement table. The others — 26 and 30 June 2014, 2 July 2010, 20 June 2008, 2 February 2018 — are days no pinned document names. What is claimed: the printed number is not the methodology’s value for the printed rate on the pinned calendar under the pre-registered settlements. What is not claimed: that the publisher erred — a settlement rule or a calendar the publisher used and did not publish is a NEEDS DATA reading, and the pre-registration names it as such.